/ WIT OS / Sentinel
Runtime security for AI agents.
AI agents introduce a new attack surface: prompts, tools, outputs, and identities. Sentinel is the runtime layer that secures them — DLP, prompt-injection defense, output validation, and a circuit breaker that stops bad days before they start.
- MITRE ATLAS-aligned threat model for AI systems
- Inline DLP across prompts, tools, and outputs
- Prompt-injection and jailbreak detection
- Circuit breaker auto-isolates compromised agents
/ What you get
Everything you need. Nothing you don't.
MITRE ATLAS coverage
Detection and response mapped to the full ATLAS matrix — prompt injection, model evasion, supply chain, exfiltration.
Inline DLP
PII, PHI, secrets, source code — detected and redacted in prompts, tool calls, and model outputs in real time.
Prompt-injection defense
Indirect injection, context smuggling, and jailbreak attempts blocked before reaching the model.
Output validation
Every model output is type-checked, schema-validated, and policy-evaluated before it leaves the agent.
Circuit breaker
Misbehaving agent? Sentinel isolates it from tools and humans within milliseconds and pages the on-call.
Full audit trail
Every prompt, tool call, output, and policy decision is logged immutably. Forensics in minutes, not weeks.
Ready to run on WIT OS?
Talk to the team about a managed deployment, a pilot, or a custom agent — we typically respond within an hour.
/ FAQ
Frequently asked questions
How does Sentinel prevent prompt injection?
Sentinel sits between agents and tools/output. It detects prompt injection patterns (instruction smuggling, jailbreak attempts, indirect injection from retrieved documents), validates outputs against schemas, runs circuit breakers on anomalous behavior, and writes immutable audit logs. Every agent action passes through Sentinel before reaching production systems.
What's the difference between Sentinel and a generic LLM firewall?
Generic LLM firewalls inspect prompts and responses in isolation. Sentinel is context-aware — it knows the agent's role, the tool the agent is calling, the data sensitivity, and the current task objective. A generic firewall might miss data exfiltration disguised as legitimate output; Sentinel catches it because it understands the workflow context.
Can Sentinel block specific data categories from leaving an environment?
Yes. Sentinel includes DLP-style policies for PII, PHI, financial data, source code, and customer-defined patterns. Outbound text from agents is inspected before delivery and blocked, redacted, or escalated based on policy.